Electrical load forecasting using artificial neural network kohonen methode Galang Jiwo Syeto / EEPIS-ITS ITS 7406.040.058
INTRODUCTION Electricity can not be stored in a large scale, therefore this power must be provided when needed. As a result there is a problem in unfixed electrical power quota,, how to operate an electric power system that always able to meet the power demand at any time, in a good quality.
INTRODUCTION The first prerequisite should be implemented to achieve the goal,the electric company must knows the electrical load or power demand in the future. So that s s why we need, ELECTRICAL LOAD FORECASTING
Final project Objectives Build electrical load forecasting system more accurate in minimum average error Compare hybrid backpropagation kohonen and hybrid counterpropagation kohonen in electrical forecasting system
Problems How to determine algorithm using backpropagation with kohonen and counterpropagation with kohonen for electrical forecasting and get minimum error How to determine the number of hidden layer in backpropagation and counter propagation methode
Limitations Issue Used Artificial Neural Network especially backpropagation,counterpropagation and kohonen. Input data used is the electrical load data taken from PLN Company, Channeling and Load Management Center division for East Java and Bali between September,1 st 2005 until January,30 th 2006 Static input data in.txt file
System design (ANN Architecture) Hybrid methode backpropagation-kohonen kohonen 2 node in input layer,4 node in hidden layer,2 node in output layer (BP) 122 node in input layer and 122 node in output layer kohonen
System Design (General Method) First,we calculate mean and standard deviation each day Second,we calculate normalization profile each day Third, Get the prediction for the mean and standard deviation for next day Get the prediction
System design (ANN Architecture) Hybrid methode counterpropagation-kohonen 2 node in input layer,4 node in hidden layer,2 node in output layer (CP) 122 node in input layer and 122 node in output layer kohonen
System Design (Data) Data for this system was electrical load data taken from PLN Company, Channeling and Load Management Center division for East Java and Bali between September,1 st 2005 until January,30 th 2006 6 per hour in mega-watt units,total of the data are 3648 data
System Design (Data) PERIODE I NPUT PERIODE O UTPUT DATA TRAINING DATA TES
Implementation (BP initialization) Initial 3 weight randomize between 0 until 1 Initial Alpha for backpropagation Maximum error value Sigmoid function value (lambda) Initial Alpha for kohonen Epoch value for kohonen
Implementation (CP initialization) Initial 3 weight randomize between 0 until 1 Initial Learning Rate alpha and beta width neighbors controller Funct ction (k0), Number of counterpropagation Epoch Initial Alpha for kohonen Epoch value for kohonen
Implementation (data preprocessing) Normalization Mean and deviation standart
Implementation (data preprocessing) Normalization Profil
Implementation (NN Training) Backpropagation Counterpropagation Get the best weight for the mean and standar deviation forecasting
Implementation (Classification) Classify mean and deviation standart in 122 group classification Classify mean and deviation standart result of the forecasting Get the normalization profil index
Implementation (Postprocessing) Use euclidiance distance to get the nearest normalization profil index
Implementation (Forecasting Result) Get the forecasting result using
Implementation (Mean Square Error) To assess the performance of this forecasting system, we use MSE for getting error calculation MSE = (Actual i - Fitted i ) 2 /n
Testing and Analyzation Determine the number of neuron in hidden layer (backpropagation-kohonen) kohonen) Jumlah Neuron 2 3 4 5 6 MSE Training 0,000007 0,000012 0,000021 0,000006 0,000002 MSE Peramalan 239242,8067 235178,5398 230566,4836 240897,7882 248966,8643
Testing and Analyzation Determine the number of neuron in hidden layer (counterpropagation-kohonen) Jumlah Neuron 2 3 4 5 6 MSE Training 0,013299 0,069946 0,094297 0,078252 0,079253 MSE Peramalan 382313,4391 451102,7709 2010904,1890 1714283,0143 10418776,1022
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen One next step forecasting metode CPNN-Kohonen BPNN-Kohonen MSE 406242,4146 893066,4288
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen Five next step forecasting metode MSE CPNN-Kohonen BPNN-Kohonen 418774,6333 342792,2535
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen Ten next step forecasting metode CPNN-Kohonen BPNN-Kohonen MSE 305903,4858 253453,5660
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen Fifteen next step forecasting metode CPNN-Kohonen BPNN-Kohonen MSE 329035,1536 249370,5583
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen Twenty next step forecasting metode CPNN-Kohonen BPNN-Kohonen MSE 396489,3527 265417,9772
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen Twenty five next step forecasting metode CPNN-Kohonen BPNN-Kohonen MSE 395924,4824 260922,8640
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen Thirty next step forecasting metode CPNN-Kohonen BPNN-Kohonen MSE 382313,4391 248966,8463
Testing and Analyzation Compare forecasting value hybrid from backpropagation kohonen and counterpropagation kohonen 1000000 mse 900000 800000 700000 600000 500000 400000 300000 200000 100000 0 1 5 10 15 20 25 30 BPNN-KOHONEN CPNN-KOHONEN tahap peramalan
Conclusion 1. Electrical load forecasting using hybrid backpropagation kohonen better than counterpropagation kohonen 2. Forecasting system that hybrid Counterpropagation Kohonen, at the training process,, if the number of neurons used in the hidden layer increase,give effect increasing the error in forecasting result. 3. Results of a small error when the training has not given effect produce a small error value at the end of the forecast. 4. For get the best result of forecasting depend on number of hidden layer that used when training process